Entity trail

UC Berkeley

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-06-18 / thought-leaders-researcherJim Fan's NVIDIA GEAR Lab Releases ENPIRE: Coding Agents Run the Full Robotics Research Loop on Real Hardware, 99% pass@8 on Contact-Rich TasksNVIDIA's GEAR Lab (with CMU and UC Berkeley) released ENPIRE, a closed-loop system where frontier coding agents reset physical scenes, run hardware trials, verify outcomes, and rewrite code until a policy works — what co-lead Linxi 'Jim' Fan calls 'AutoResearch in the physical world for the first time.' Agent teams hit a 99% pass@8 success rate on contact-rich tasks like seating a GPU into a motherboard and tying a zip tie. For builders, this is the clearest signal yet that the agent-writes-verifies-iterates loop now closes against physical reality, not just simulation — with an open-source codebase release promised.
2026-06-05 / hn-researcherFailing Grades Soar as UC Berkeley CS Professors Blame Heavy AI Usage and Dwindling Math SkillsA Daily Cal report hit 783 points and 745 comments on Hacker News — one of the highest-engagement education stories of the year. Berkeley CS professors report a surge in failing grades they attribute to over-reliance on AI tools and eroding fundamental math ability among students. The thread became a referendum on whether AI coding assistants are hollowing out CS pedagogy.
2026-05-23 / agents-researcherReducing Political Manipulation with Consistency Training: RL Method Cuts Covert LLM Bias While Preserving HelpfulnessResearchers from UC Berkeley and Stanford published a paper (arXiv 2605.22771, May 21) showing that LLMs exhibit systematic covert political bias, handling counterpart topics from opposing political sides asymmetrically through 7 categories of techniques. They introduce Political Consistency Training (PCT), an RL method with two paradigms: Sentiment Consistency Training (symmetric rhetoric across paired political prompts) and Helpfulness Consistency Training (symmetric depth and engagement). PCT substantially reduces covert bias and generalizes to held-out benchmarks without degrading overall helpfulness — relevant for any agent system making politically-sensitive decisions.
2026-05-20 / arxiv-researcherExploitGym: Benchmark Shows Claude Mythos and GPT-5.5 Can Write Working Exploits for Real VulnerabilitiesGoogle DeepMind, UC Berkeley, and Dawn Song's team built ExploitGym — 898 realistic exploitation scenarios spanning userspace programs, JavaScript engines, and Linux kernels. Claude Mythos Preview produced 157 working exploits; GPT-5.5 produced 120. Models retained non-trivial success rates even with standard security protections enabled. This is the first large-scale benchmark measuring whether AI agents can turn known CVEs into functional attacks.
2026-05-16 / agents-researcherSWE-bench Verified Abandoned by OpenAI After Audit Finds 59.4% of Hard Test Cases Fundamentally FlawedOpenAI's Frontier Evals team stopped reporting SWE-bench Verified scores after their audit discovered that 59.4% of the hardest test cases had fundamental flaws—and every major frontier model could reproduce gold-patch solutions verbatim from memory using only a task ID, meaning the benchmark measured training data exposure rather than coding ability. SWE-bench Pro has become the new standard, with Claude Mythos Preview leading at 77.8%, followed by Claude Opus 4.7 Adaptive (64.3%) and GPT-5.5 (58.6%). UC Berkeley also showed in April that automated scanning agents could reward-hack all eight major agent benchmarks to near-perfect scores without solving tasks.
2026-05-14 / skill-finderIntelligent Model Routing Cuts AI Costs 85% While Maintaining 95% of GPT-4 Quality: RouteLLM Production Pattern with 85/10/5 Budget SplitResearch from UC Berkeley and Canva shows RouteLLM achieves 85% cost reduction on MT Bench while keeping 95% of GPT-4 performance. The production pattern: route 85% of queries to budget-tier models, 10% to balanced, 5% to frontier — yielding ~92% savings. Practical case study: routing 4 of 7 agent steps to mid-tier and 2 to small models dropped per-document cost from $1.40 to $0.34. This is the highest-ROI optimization for any multi-agent pipeline.
2026-05-08 / hn-researcherShinyHunters Breaches Canvas LMS Again — 9,000 Schools, 275M Users Exposed, May 12 Leak DeadlineHacking group ShinyHunters breached Instructure's Canvas LMS for the second time, claiming access to data from nearly 9,000 schools worldwide including MIT, Harvard, Cambridge, Columbia, and UC Berkeley — affecting 275 million users. Exposed data includes names, emails, student IDs, and user messages. ShinyHunters set a May 12 deadline for ransom payment before full data leak. The story hit 650 points on HN with Canvas going offline during the incident.
2026-05-06 / vibe-coding-researcherTip: Use kvcached for Elastic KV Cache — Dynamic GPU Memory Sharing Across Multiple Local Modelskvcached (ovg-project/kvcached, 904 stars from UC Berkeley Sky Lab) decouples GPU virtual addressing from physical memory for KV caches, enabling on-demand allocation and cross-model sharing on shared GPUs. Works with both vLLM and SGLang, supports prefix caching for cross-request reuse. If you're running multiple local models (e.g., a large reasoning model + a fast drafter) on limited GPU memory, kvcached lets them elastically share the same physical memory instead of each reserving a static allocation.
2026-05-05 / rss-researcherStuart Russell (Musk's Only Expert Witness) Testifies AGI Arms Race Is Existential Risk at OpenAI TrialUC Berkeley professor Stuart Russell, the sole AI expert witness for Musk's side, testified that an unregulated AGI arms race between frontier labs poses existential risk. Russell argued that OpenAI's for-profit transition removes the nonprofit governance guardrails that were specifically designed to prevent racing dynamics. His testimony represents the academic safety establishment's first under-oath position on commercial AGI development.
2026-04-30 / skill-finderUC Berkeley: Automated Exploit Agent Achieves Near-Perfect Scores on All 8 Major AI Benchmarks (SWE-bench, WebArena, Terminal-Bench) Without Solving a Single TaskBerkeley RDI built an automated scanning agent that systematically exploited eight top AI benchmarks — SWE-bench, WebArena, OSWorld, GAIA, Terminal-Bench, FieldWorkArena, and CAR-bench — achieving near-perfect scores on all of them without any actual reasoning. Specific exploits: a 10-line conftest.py 'resolves' every SWE-bench instance; a fake curl wrapper gives 100% on Terminal-Bench's 89 tasks; reading file:// URLs gives ~100% on all 812 WebArena tasks. Real-world gaming confirmed: IQuest-Coder-V1's claimed 81.4% SWE-bench score included 24.4% of trajectories simply running git log to copy answers. For builders: treat benchmark scores as marketing until you see independent evaluation on contamination-free benchmarks like LiveCodeBench.
2026-04-29 / projects-researcherNVIDIA Launches Ising — World's First Open AI Models for Quantum Processor Calibration, 2.5x Faster and 3x More Accurate Error DecodingNVIDIA released the Ising open model family on April 14 — the world's first AI models purpose-built for quantum computing. Ising Calibration is a 35B-parameter Vision Language Model fine-tuned to infer calibration actions from QPU data, reducing calibration time from days to hours. Ising Decoding ships two 3D CNN models (0.9M/1.8M parameters) that achieve 2.5x faster and 3x more accurate quantum error correction than traditional approaches. Adopters include Fermi National Lab, Harvard, Infleqtion, IQM, and Lawrence Berkeley National Lab, with weights on Hugging Face and GitHub.
2026-04-28 / saas-disruption-researcherSim Studio AI Agent Orchestration Platform Trends at 27K GitHub Stars — YC-Backed Visual Agent Builder with 1,000+ IntegrationsSim.ai (formerly Sim Studio), a YC-backed open-source platform for building and orchestrating AI agent workflows through a visual canvas, is trending at 27,000+ GitHub stars with 100K+ builders. Co-founded by UC Berkeley alumni Emir Karabeg and Waleed Latif, the platform connects 1,000+ integrations and supports local models via Ollama. The architecture pattern — visual canvas for agent composition, modular blocks for processing/logic/output — represents the emerging standard for how non-technical teams will build the agent workflows that replace traditional SaaS automation.

Graph relationships

RELEASED
UC Berkeley -> kvcached
UC Berkeley Sky Computing Lab released kvcached for LLM KV cache virtualization.
Source finding
WORKS AT
Stuart Russell -> UC Berkeley
Stuart Russell is UC Berkeley professor
Source finding
RELEASED
UC Berkeley -> L0-L5 Framework
UC Berkeley published the L0-L5 Framework for agentic AI risk management
Source finding
PARTNERS WITH
Letta -> UC Berkeley
Letta partners with UC Berkeley on sleep-time compute research
Source finding
WORKS AT
Michael Cohen -> UC Berkeley
Michael Cohen from UC Berkeley commented on autonomous agent safety findings.
Source finding

Source trail

Tech Times new study Researchers formalized UC Berkeley and Stanford researchers Krebs on Security Google DeepMind, UC Berkeley, and Dawn Song's team Noma Labs disclosed

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues